Codex is turning AI coding into a mobile approval loop
The interesting Codex news today is not that someone can "code from a phone."
It is that the phone is becoming a review surface.
OpenAI's post on working with Codex from anywhere puts Codex inside the ChatGPT mobile app for starting work, watching tasks, steering execution, reviewing results, and approving next steps while the actual job keeps running on a laptop, Mac mini, or devbox.
That changes the feel of the workflow. The agent stays close to the codebase. The human does not have to stay close to the machine.
For small SaaS teams and solo builders, that matters. A useful agent workflow is rarely "make the model smarter and hope." It is usually much more practical: give the agent a bounded task, let it do the mechanical work, then make approval easy at the moment judgment is needed.
The story
Many AI coding demos still assume the developer is sitting there watching the terminal. The agent opens files, edits code, runs checks, and waits for the next yes or no.
Mobile Codex changes that posture. A founder could start a small docs patch before lunch, glance at the diff from a phone, approve the next step away from the desk, and return later to the branch. The work still happens in the right environment. The human review just becomes easier to reach.
That may be the real product move. Codex is being pushed from a terminal session toward a distributed approval workflow.
OpenAI also published Sea Limited's perspective on agentic software development with Codex. It is a customer story, so read it with that filter. Still, the positioning is clear: OpenAI wants Codex to look credible for teams, not just solo experiments.
What actually changed
The clean headline is Codex in the ChatGPT mobile app. The useful version for builders is review from anywhere, not pretending a phone is a better IDE.
Underneath that, the Codex CLI alpha channel kept moving with 0.131.0 prereleases: alpha.16, alpha.18, and alpha.19. These are prerelease builds with thin release notes, best treated as sandbox material. The takeaway is project velocity, not a production upgrade recommendation.
The more useful Codex CLI commits are around the parts teams notice once agents touch real work: signed macOS release promotion, workspace root summaries, runtime workspace roots, and profile identity constraints. Those are not launch-video features. They are the boundaries that help answer a harder question: what can this agent touch, and as whom?
Hermes Agent had a similar plumbing day. A PyPI publish workflow, uvx-based ACP registry distribution, and Zed ACP metadata make the project easier to package and wire into editor-style agent clients.
Hermes also picked up smaller terminal fixes for scrollback duplication and light-mode visibility, plus a goal-judge change that raises max_tokens from 200 to 4096 and makes the value configurable. That is unglamorous work, but long agent runs need readable state and enough room to judge what happened.
Why builders should care
The product lesson is simple: the handoff is becoming part of the product.
A coding agent that edits files is useful. A coding agent that edits files inside the right workspace, explains what changed, and waits at clear approval points is easier to trust on the second run.
If you are building with agents, this is the part to copy. Do not only ask whether the model can finish the task. Ask where the agent should stop, what the human needs to see, and how approval gets captured without turning every run into babysitting.
For small teams, the right first targets are still boring: docs patches, typo fixes, test updates, issue triage, QA passes, and narrow refactors with checks. The agent can do the repetitive work. The human keeps the judgment call.
What to try today
- Use Codex mobile on one low-risk task where review matters more than deep architecture.
- Keep the Codex CLI
0.131.0alpha releases in a sandbox. They are prerelease builds, not a production baseline. - Mark the approval points in your own agent workflow. Decide where the agent should pause before it changes more context.
- Treat workspace boundaries, install paths, logs, and profile identity as product features if you are building developer tools.
- Turn one internal agent run into a short tutorial that shows the task, approval step, verification, and final artifact.
Builder takeaway
Today was not really a stable-release story. It was a workflow story.
Codex is moving review and approval onto the phone. Codex CLI and Hermes Agent are tightening the runtime underneath. The practical move is to stop treating agents as one-off terminal sessions and start designing the handoff around them.
Sources checked: Hermes Agent releases/commits, OpenAI Codex releases/commits, OpenAI RSS/news | Window: 24h | Generated: 2026-05-15